It’s fair to say DeepSeek has arrived.

The Chinese startup and its R1 model exploded onto the AI scene last week, and - at least temporarily - turned the industry on its head.

With OpenAI having established itself as the market leader over the last few years, DeepSeek’s sudden and massive hype appears to represent the most serious threat to its dominance to date.

In this article, we’ll take a look at why there’s so much excitement about DeepSeek R1 and how it stacks up against OpenAI o1.

The Interface

The very first thing you’ll notice when you open up DeepSeek chat window is it basically looks exactly the same as the ChatGPT interface, with some slight tweaks in the colour scheme.

This is, frankly speaking, a good move by the DeepSeek team. While the rights-and-wrongs of essentially copying another website’s UI are debatable, by using a layout and UI elements ChatGPT users are familiar with, DeepSeek reduces friction and lowers the on-ramp for new users to get started with it.

Thinking Through Problems

What surprised many R1 was released was that it included the thought-process feature present in OpenAI’s o1 model. This means that users can now see how the model arrived at a particular conclusion by reading the log of its thought-process, otherwise known as the chain of thoughts.

While this puts the two on an even keel in terms of the transparency with which they work through problems, it’ll be interesting to see if any future studies look more closely at the actual quality of thinking that the models do.

Logic and Reasoning

ChatGPT is far from perfect when it comes to logic and reasoning, and like any model its prone to hallucinating and stubbonly instisting it is correct when it is not.

However, much to the surprise of many given how advanced ChatGPT’s model appear, DeepSeek’s R1 performs better than o1 in most aspects related to logic, reasoning, coding and mathematics.

It makes DeepSeek a clear winner in this domain, and one that will help it carve out its place in the market, likely becoming more popular with engineers, programmers, mathemeticians and STEM related roles as the word gets out.

Creativity

Many of the analyses done on LLM models focus almost exclusively on technical aspects like network response times in order to measure the differences between the models, rather than the broader cognitive abilities the LLM is capable of demonstrating.

While it might have strengths with regards to logical thinking, it quite simply lacks the features to demonstrate the full range of its abilities.

Most notably, R1 is missing the ability to generate images, meaning that while it might enable creativity, the type of creativity that it enables is limited, compared to o1.

Interactivity and User Experience

So we’ve discussed the similarities in the user interface, but we’ve also identified the fact that features are missing that are quite important in terms of how users use them.

As well as the image-generation we mentioned before, DeepSeek does not offer voice mode, which aside from being an accessibility feature, is a handy way to engage with the tool.

Furthermore, DeepSeek does not give users the ability to edit responses. OpenAI offers Canvas, which lets users work with ChatGPT responses like a live document, making it easier to use as a springboard for ideas. OpenAI takes this one.

Cost

This might be the only category for which there is a relatively clear winner, and it is in some ways the reason that DeepSeek caused such a stir when it opened the gates on its R1 model.

It has been widely reported that Bernstein tech analysts estimated that the cost of R1 per token was 96% lower than OpenAI’s o1 reasoning model, but the root source for this is surprisingly difficult to find.

DeepSeek themselves say it took only $6 million to train its model, a number representing around 3-5% of what OpenAI spent to each the same goal, though this figure has been called wildly inaccurate.

DeepSeek goes further than this, to say on its official WeChat account that R1 is “20 to 50 times cheaper to use” than ChatGPT’s, with its API costing $0.14 for one million tokens or roughly 750,000 words, compard to $7.50 per one million ChatGPT tokens.

So, it appears that some of these claims have been (surprise!) exaggerated in the name of marketing, but are likely to point to some kind of truth.

The announcements impacting the market so much that NVIDIA’s share prices plummeted (with a record one-day loss), with the developments suggesting their dominance - and that of OpenAI - might be in jeapoardy.

Biases

ChatGPT’s biases are clear and numerous. It has a Western view of the world that OpenAI ask users to remember when using it, and all of the models have revealed clear issues with how data is indexed, interpreted and then ultimately sent back to the end-user.

While we may not know as much just yet about how DeepSeek R1’s biases impact the results it will give, it has already been noted that its results have strong slants, particularly the ones given to users in China, where results will parrot the views of the Chinese Communist Party.

Privacy

Much like the previous section, your opinion on the way that users privacy is handled by OpenAI or by DeepSeek will very much rely on your location, political views and wider worldwide.

If you are in the West, you might be concerned about the way that Chinese companies like DeepSeek are accessing, storing and using the data of its users around the world. Inversely, users living in the East are likely to have similar concerns about OpenAI for the same reasons.

Users are right to be concerned about this, in all directions.

These tools have become wildly popular and with users giving huge amounts of data to them it is only right that this is treat with a strong degree of skepticism.

Conclusion

In conclusion, the arrival of DeepSeek’s R1 model has undoubtedly shaken the generative AI landscape, offering a formidable challenge to OpenAI’s dominance.

While it boasts notable strengths, particularly in logical reasoning, coding, and mathematics, it also highlights significant limitations, such as a lack of creativity-focused features like image generation. The cost advantage of DeepSeek R1 is a major selling point, though some of the marketing claims may be exaggerated.

Additionally, issues like bias and privacy concerns remain central to the debate around both models, with geopolitical perspectives influencing opinions on data handling. Ultimately, while DeepSeek R1 presents a promising alternative, whether it will truly disrupt the market in the long term remains to be seen.